How it works

What this site does

sportsanalysis.info turns public NFL and college football data into answers: who covered, who went over, which lines moved, and how each game shapes up. Every week has a game list with a game card for each game, every team has a page with its schedule, lines and ATS splits, and Ask SportsAnalysis answers questions in plain English.

How the model works

• Power rating: an opponent-adjusted scoring margin (SRS-style). NFL margins are capped to limit blowouts (college margins count in full, and every FCS opponent has its own rating), home field is removed, and each season starts from last season’s rating regressed toward average, so early numbers are stable.
• Beyond the score: each game’s points are blended with the points its yards and turnovers usually produce (turnovers are mostly luck, so they count less); the preseason starting point also leans on the ratings implied by this season’s closing lines; and where the backtest showed a gain, a backup starting quarterback (NFL) and strong wind at kickoff adjust the line and total. The rating settings are chosen by a walk-forward backtest over every season we hold.
• Prediction: home margin = rating(home) − rating(away) + home field (none at neutral sites), blended with the market spread. Win probability comes from that margin; projected scores from offense and defense ratings.
• Live engine: the projections and leans on this site come from the engine live for each league; today that is the power ratings in both leagues. XGBoost, a learned correction to the closing line built on top of these ratings, is retrained every night and compared with them and with the line on Deep Dive; either engine can be made live per league.
• Leans: an ATS lean when the model and market disagree by 2.5–7 points (NFL) or 3–10 (college); an over/under lean when the model total is 5+ points (NFL) or 5+ (college) from the line. Leans become plays only once their record is a proven edge.
• Game-card signals (trends, line movement) are tested separately; only ones that beat the market in the backtest turn green.

Walk-forward backtest over the 2016–26 seasons: every week is predicted using only games played before it. Break-even against the spread at standard −110 odds is 52.4%; a record only counts as a proven edge with 40+ decisions and a win rate above break-even by more than luck explains (one-sided z ≥ 1.64). Until then, model leans are shown as context, not plays. Beat the closing line is how often the line moved toward the model’s side after it opened: the quickest honest test of whether the model knows something the market doesn’t.

Walk-forward backtest over the 2016–26 seasons: every week is predicted using only games played before it. Break-even against the spread at standard −110 odds is 52.4%; a record only counts as a proven edge with 40+ decisions and a win rate above break-even by more than luck explains (one-sided z ≥ 1.64). Until then, model leans are shown as context, not plays. Beat the closing line is how often the line moved toward the model’s side after it opened: the quickest honest test of whether the model knows something the market doesn’t.

• Scores, schedules and lines: ESPN’s public scoreboard and game APIs (lines as published there by its sportsbook partner), synced every 15 minutes.
• Opening vs current lines: ESPN odds history, refreshed every 15 minutes for games in the next 10 days.
• Injury reports: ESPN (NFL); college teams don’t publish them.
• Box scores (yards, turnovers, starting quarterbacks): ESPN game summaries, back to 2015.
• Kickoff weather: Weather data by Open-Meteo.com (CC BY 4.0): historical observations for past games, forecasts for upcoming ones.
• News & sentiment: ESPN and Google News headlines, scored with a simple keyword method.
• Ratings and game cards are rebuilt every 10 minutes. Seasons: current plus the three before it.
• Ask SportsAnalysis uses an AI model only to understand the question; every number comes from the data above.